Three weeks ago, we released the first version of Edge0: “A 35B language model running on a local device, using only 1–2.5 GB of peak memory.” Today, we’re taking the next step. We’re open-sourcing the Edge0 app layer—now available on: 📱 iOS 🖥️ macOS 🤖 Android 🪟 Windows One... app. Four platforms. Fully on-device.show more

The Edge0 App layer is now open source: 🔗 This release includes: 📱 Cross-platform app code for iOS, macOS, Android, and Windows 🧠 Local model + runtime integration 🗂️ Model management 🛠️ Build and deployment guides Everything you need to run Edge0 locally and start building on top of it.

The first Edge0 release proved that large models can run on consumer devices. This release takes the next step: From an on-device model to an on-device AI application. Developers can now build local AI assistants, creative tools, smart hardware experiences, and more—across phones, computers, and other devices. Our direction is simple: ☁️ Less dependence on the cloud. 🧠 More intelligence on the device. 🎛️ More control for developers and users.

At its core, Edge0 keeps the full model in local storage and loads only what each step needs. 🧠 Current experts are loaded into memory on demand. 🔮 Next experts are predicted and prepared in advance. ⚡ Storage reads run in parallel with model computation. ♻️ Memory is reused instead of holding the entire model at once. 🧩 LoRA adapters run alongside the base model without modifying it. The result: A large model with a small active working set, running fully on-device. Try Edge0, build with it, and share what you create.

What’s next for Edge0? Our Q4 2026 roadmap includes: 🧩 A unified inference framework with one API across iOS, macOS, Android, Windows, and Python 🔄 Runtime adaptation across hardware platforms 🧠 A CUDA backend for the Python framework 🧬 Support for next-generation model architectures ⚡ Latent thinking and batched expert prediction to reduce end-to-end reasoning latency 🛠️ More model tiers and adapter releases The goal is simple: One open framework for running powerful AI locally across devices and platforms. 🔗

No love for (desktop) Linux?

Linux support is coming soon. Stay tuned!

This makes me very happy! Thank you

Do you hate Linux or something? 😭

Linux support is coming soon. Stay tuned!

Lovely. But why Do you guys ignore Linux in the age of AI? Before, it used to be windows that was ignored!

Linux support is coming soon. Stay tuned!

Yawn...come back to me when you've got a first class Linux client.

Linux support is coming soon. Stay tuned!

That's really cool, putting 35B LLM fully on-device across all four platforms, really impressive progress.

How much of the app layer is shared across all four platforms?

Today, the four platform apps still use some platform-specific code, while sharing the same core model and runtime concepts. We’re working toward a unified access layer so more of the app experience can be shared across all platforms.

Got it. Do the platform-specific pieces mostly sit at the UI layer, or do they reach into the model/runtime integration too?

It reaches into the model and runtime integration as well. The UI layer is platform-specific, but each platform also uses its native runtime stack — MLX/Metal on Apple platforms, and patched llama.cpp with CPU/Vulkan backends on Android and Windows. What we share today is the model/runtime contract and the core Edge0 inference approach. We’re working toward a unified access layer across all four platforms.

That’s more of the stack than I expected, especially separate MLX/Metal and llama.cpp paths. Is the unified layer mainly standardizing model loading and inference calls, or will it eventually replace those backend-specific runtimes?

Everything except for Linux. 😭

Linux support is coming soon. Stay tuned!

35B in 1–2.5 GB still sounds slightly illegal. 😑 interesting... I'll try later 😎

35B at 1-2.5 GB peak is the number that matters. open-sourcing the app layer on top of it - curious what the actual cold-start looks like on mobile before I believe the 'instant' part.

It's awesome that the same code base runs on iOS, macOS, Android, and Windows. It really makes local AI seem a lot more significant.

Thanks! That’s exactly what we’re aiming for. A consistent local AI experience across iOS, macOS, Android, and Windows makes on-device AI much easier to use and build on.

¿por qué no GNU/Linux?

Linux support is coming soon. Stay tuned!

¡ seré tu betatester !

i think people will feel stupid for not curating their own knowledge base once models like this run locally. the model is the easy bit, the notes are what take years

Absolutely. Once capable models run locally, the real advantage comes from the context and knowledge people build around them. Local AI makes that knowledge more private, persistent, and useful.

yh and the part i care about is the notes being plain files, so whichever local model wins next year can just read them

Linux?

Linux support is coming soon. Stay tuned!

What's first-token latency like on older phones? The streaming approach has me curious.

I think you need to fix your GitHub repo,its not working on windows

@grok no entiendo, explicame

有直接安装可用的apk吗？

@grok what is this app

You should make a TestFlight for iOS

8b does'nt looks good

Thanks for flagging this! That output doesn’t look right—it may be related to the build process. Could you open a GitHub issue with your hardware specs, build steps, and logs so we can investigate?

Do you have benchmarks if Edge0 against others models ?

How good is the model? In different use cases?

Woooh! Congrats guys :)

四端共用一套权重还能输出一致才是硬骨头

是的，我们下一个版本就是开源统一推理引擎

Linux . . . . . . ?

Linux support is coming soon. Stay tuned!

The app layer is open. Now we want to see what you build. 🛠️ What's the first thing you'd run with a 35B model that never touches the cloud? Best ideas get a shoutout from us 👇

35B at 1–2.5 GB peak with on-demand experts and next-expert prefetch is the Edge0 claim that matters for Mac. On the MLX/Metal path, what's cold vs warm penalty when the expert miss rate climbs?

what is the accuracy of the edge0 models